llama.cpp Adds Multi-Token Prediction for Qwen3 Family, Boosting Local Inference Speed

llama.cpp now supports MTP heads for the Qwen3 family, with Hugging Face GGUF listings flagging MTP-capable models explicitly. MTP allows the model to predict multiple tokens per forward pass, and the maintainers describe the performance jump on commodity hardware as substantial.